[CUDA] Add MatMulBlockQuantizedFp4Weight contrib operator

[tianleiwu]

Description

Adds a new com.microsoft CUDA contrib operator MatMulBlockQuantizedFp8Weight : a
weight-only NVFP4 (E2M1) matrix multiplication that computes
Y = A * dequant(B)^T (+ bias).

The weight tensor B is stored as packed NVFP4 (two E2M1 values per byte, low
nibble first). The dequantized weight is
e2m1(B) * weight_scale_2 * e4m3(weight_scale[n, k / block_size]), where
weight_scale holds one E4M3 scale per block_size (default 16) consecutive K
values and weight_scale_2 is a single global fp32 scale. The weight is
dequantized to the activation type (FP16/BF16) and multiplied with the FP16/BF16
activation.

Operator schema

• Inputs: A (FP16/BF16 [..., K]), B (packed uint8 NVFP4 [N, K/2]),
  weight_scale (uint8 E4M3 [N, ceil(K/block_size)]), weight_scale_2
  (scalar fp32), optional input_scale (scalar fp32), optional bias ([N]).
• Attributes: K, N, block_size (default 16).
• Output: Y ([..., N]) in the activation type.

Implementation

• Architecture-independent dequant + GEMM/GEMV path that runs on Hopper (SM90)
  and Blackwell.
• Native SM120 CUTLASS NVFP4 path (guarded by ORT_ENABLE_BLOCKQUANT_SM120).
• ONNX schema, kernel registration, and CMake build wiring for the CUDA EP and
  the CUDA plugin EP.

Motivation and Context

Enables weight-only NVFP4 quantized matmul for LLM inference, reducing weight
memory footprint while keeping FP16/BF16 activation precision.

Testing

• Added C++ unit tests (matmul_block_scaled_fp4_test.cc) covering FP16 and
  BF16 activations, with and without bias.
• Added a Python profiling/verification script
  (profile_matmul_block_scaled.py).
• Verified the CUDA EP test binary (onnxruntime_provider_test) builds cleanly
  with ENABLE_FP4 for SM86 + SM120a.

[Copilot]

Pull request overview

This PR adds a new CUDA contrib operator com.microsoft::MatMulBlockScaledFp4 to ONNX Runtime, implementing weight-only NVFP4 (E2M1) matmul with per-block E4M3 scales and an optional opt-in native SM120 (Blackwell) CUTLASS path. It also adds C++ tests, profiling tooling, documentation, and CMake wiring for both the CUDA EP and CUDA plugin EP.

Changes:

• Introduce MatMulBlockScaledFp4 schema + CUDA kernel implementation (dequant+cuBLAS, small-M GEMV, optional native SM120 path).
• Register the operator in the MS opset and CUDA contrib kernel registry.
• Add unit tests, a Python profiling harness, and operator documentation/experiments notes.

Reviewed changes

Copilot reviewed 14 out of 14 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
onnxruntime/test/python/contrib_ops/profile_matmul_block_scaled.py	Opt-in Python accuracy/latency harness for the new CUDA contrib op.
onnxruntime/test/contrib_ops/matmul_block_scaled_fp4_test.cc	New CUDA-focused unit tests for FP16/BF16, bias, and GEMV decode path.
onnxruntime/core/graph/contrib_ops/ms_opset.h	Adds the operator to the Microsoft opset v1 schema list.
onnxruntime/core/graph/contrib_ops/contrib_defs.cc	Defines the MatMulBlockScaledFp4 schema and shape inference.
onnxruntime/contrib_ops/cuda/math/matmul_block_scaled_fp4.h	Declares the CUDA kernel class and launcher APIs.
onnxruntime/contrib_ops/cuda/math/matmul_block_scaled_fp4.cu	Implements NVFP4 dequant, bias add, and small-M fused GEMV kernels.
onnxruntime/contrib_ops/cuda/math/matmul_block_scaled_fp4.cc	Implements kernel registration, validation, dispatch, and native-path prepacking.
onnxruntime/contrib_ops/cuda/math/matmul_block_scaled_fp4_sm120.cu	Implements the native SM120 CUTLASS NVFP4 x NVFP4 GEMM path.
onnxruntime/contrib_ops/cuda/cuda_contrib_kernels.cc	Registers the new CUDA contrib kernel.
docs/contrib_ops/cuda/matmul_block_scaled_fp4.md	Adds operator documentation (format, dispatch chain, env vars, workflows).
docs/contrib_ops/cuda/matmul_block_scaled_fp4_experiments.md	Records performance experiments and rationale for rejected variants.
cmake/onnxruntime_providers_cuda.cmake	Enables SM120 build define needed for native path in the CUDA EP target(s).
cmake/onnxruntime_providers_cuda_plugin.cmake	Enables the same SM120 build define for the CUDA plugin EP target.
cmake/onnxruntime_cuda_source_filters.cmake	Ensures the SM120-specific .cu file is compiled only when arch 120 is enabled.

[github-actions]

You can commit the suggested changes from lintrunner.